Skip to content
Drix10 Blog

Qwen3.5-0.8B weight edit improves math 6x

, 8 items in AI and Robotics Applications, 4 min read

In this digest (8 items)

Qwen3.5-0.8B weights were edited by GPT-6 Astra, changing 0.016% of parameters and boosting math performance up to 6×. The edit also raised instruction-following ability up to 3×, and it was done without gradients or training.

Key points

  • Edit scope: only 0.016% of the model’s weights were changed.

  • Performance gain: math reasoning improved up to 6×; instruction following improved up to 3×.

  • Method: the changes were made without using gradients or any training process.

Sources

Qwen3.5-0.8B Weight-Space Programming improves math 6×, instruction following 3×

Qwen3.5-0.8B was edited by GPT-6 Astra using Weight-Space Programming, changing only 0.016% of weights and boosting math performance up to 6× and instruction following up to 3×, without gradients or training.

Key points

  • Changed only 0.016% of Qwen3.5-0.8B weights.

  • Improved math up to 6× and instruction following 3×, without gradients or training.

Sources

3D World Simulator launches with 50+ FPS RTX 5090 performance

3D World Simulator is a multimodal world model that runs at over 50 frames per second on an RTX 5090. It predicts a unified 3D representation that can generate novel view appearance, multi-view consistent RGB-D, 3D geometry, and spatially aligned tactile signals. An interactive demo streams renders to a browser and supports robot arm control. The system uses a hybrid particle-voxel autoencoder and a 3D U-Net dynamics model.

Key points

  • 50+ FPS on RTX 5090

  • Generates novel view appearance, multi-view RGB-D, 3D geometry, tactile signals

Sources

Tutor Intelligence raises $75M Series B, targets 1,000 robots by 2027

Tutor Intelligence announced a $75 million Series B funding round led by a16z. The company aims to have over 1,000 robots under contract by the end of 2027.

Key points

  • Funding: $75M Series B led by a16z with participation from USV and Neo.

  • Deployment goal: more than 1,000 robots under contract by end of 2027.

Sources

JEM SSL adds latent patch-level objective for 7B params

JEM is a principled SSL method derived from information theory that introduces a latent patch-level objective scaling from 300M to 7B parameters. Training is stable and it matches or beats DINOv3.

Key points

  • Derived from information theory

  • First latent patch-level objective demonstrated from 300M→7B params

Sources

Terminus Notes launched, open-source replacement for Notion

Terminus Notes is a free, open-source tool that mimics Notion while removing AI credits and sidebars. It can export data or connect to Claude.

Key points

  • 100% Free, 100% Open source, 100% Customizable.

  • Export your data or point it to your Claude at any time.

Sources

Sound and Complete built, wins first place at Vibecheck hackathon

Sound and Complete is a verified tool to help guitarists practice, announced as the winner of first place overall at the Vibecheck formal methods hackathon. The creators also note they met many people and learned from the event.

Key points

  • First place overall at the Vibecheck formal methods hackathon

  • Built Sound and Complete, a verified tool to help guitarists practice

Sources

Robotics startup certificate hurdle, target Q1 2027

Robotics startup reports that the robot’s biggest problem is a certification, not software, and aims to obtain the required certificate by Q1 2027. Design and software are completed, with 97% correct defect detection on held-out photos.

Key points

  • Certification target: Q1 2027 for ASME A120.1, OSHA 1910.66, NRTL listing.

  • Performance claim: 97% correct defect detection on held-out photos.

Sources